Multiscale Enhanced Sampling Using Machine Learning.

Area of Science:

• Computational Biology
• Molecular Dynamics
• Machine Learning

Background:

• Multiscale enhanced sampling (MSES) enhances protein structure sampling by coupling all-atom dynamics with coarse-grained (CG) models.
• Existing MSES methods rely on CG models, which may not fully capture complex protein dynamics.

Purpose of the Study:

• To develop an extended MSES approach that replaces the traditional CG model with a machine learning-based reduced subspace.
• To utilize a variational autoencoder (VAE) for generating low-dimensional representations of protein dynamics.

Main Methods:

• Trained a VAE model using molecular dynamics (MD) trajectories of the ribose-binding protein (RBP), extracting inter-residue distances as features.
• Generated interpolated structural data in the VAE's latent space, representing transitions between closed and open RBP forms.
• Decoded latent space data into time-series inter-residue distances to drive atomistic sampling via MSES.

Main Results:

• The VAE's latent space effectively characterized the distinct structural dynamics of RBP's closed and open forms.
• Generated data in the latent space was refined into distinct closed and open basins using MD simulations.
• The VAE-enhanced MSES successfully recovered the correct structural ensemble, demonstrating improved sampling efficiency.

Conclusions:

• The proposed VAE-based MSES extension offers an efficient and accurate method for enhanced protein structure sampling.
• This approach effectively captures conformational changes and improves the recovery of complex protein structural ensembles.
• Integrating machine learning with enhanced sampling techniques holds significant promise for molecular simulations.